# Method — local-agent-turn-time-budget-2026 (research stage, 2026-11-08) ## What was attempted A bounded single-stream prefill/decode measurement on the host this stage ran on. - Host: 2 logical vCPUs, 11 GB RAM, Linux 6.17.0-1019-oracle (shared sandbox, other services resident). - Backend: Ollama HTTP API, `POST http://127.0.0.1:11434/api/generate`, `stream:false`. - Model artifact: `qwen3-vl:2b`, 2.1B params, `Q4_K_M`, 1,889,519,687 bytes, digest `0635d9d857d497aeadba3d7d27485746c50554446f9f6ec01ef39788221adbe`, modified 2026-11-03T16:12:51Z (from `/api/tags`). - Request options: `num_predict: 48`, `num_ctx: 8192`, `temperature: 0`, `seed: 42`, `keep_alive: "10m"`. - Fixtures: three prompts of different length (`short` ~32 prompt tokens, `medium`, `long`), 3 repeats each. Two separate time limits applied to this run; see "Two limits" below. - Derived rates, not server-reported headline numbers: `prefill_tps = prompt_eval_count / (prompt_eval_duration_ns / 1e9)`, `decode_tps = eval_count / (eval_duration_ns / 1e9)`. ## Two limits, both on the record One run was governed by two different time limits, and the evidence keeps both: - **Declared run budget: 110 s.** This is the deadline the stage was given for the measurement (`run_cap_s` 110 in the selection record), and it is the number the article quotes. - **Script kill cap: 130 s.** The driver script this stage used was itself launched with a 130 s kill cap, a ceiling above the declared budget that ended the run once the budget was already spent and a slow request was still outstanding. The `provenance` field in both `runs.jsonl` records says exactly that: "killed by its 130 s cap before flushing runs.jsonl". Neither number replaces the other: the 110 s figure is the budget the run was meant to fit in, the 130 s figure is the hard stop that actually fired. Because the script buffered its records until the end of the loop, no record was on disk at that point; the two surviving records were recovered verbatim from the stage's own stdout and re-written with the provenance string quoted above. Their measured fields (`wall_s`, `prompt_eval_count`, `eval_count`, `prefill_tps`, `decode_tps`) are unchanged from the printed output and were not re-timed, edited or rounded in this reconciliation. ## Result: rejected as noise 2 of 9 planned runs completed inside the deadline. Observed `decode_tps` 1.615 and 3.165 (96% spread), `prefill_tps` 3.781 and 228.859 (the first run includes cold model load). Wall time for the same fixture and token counts: 79.8 s vs 16.9 s. The selection's failure condition requires prefill and decode to be separated with under 10% spread across three repeats. That gate is not met, so these numbers are **not** a baseline and must never appear as a measured tokens/second figure. No GPU was measured; no cross-hardware claim is possible from this host. ## Reproduce The fixture and script are deterministic; the variance is the host, not the harness. A rerun needs a quiet machine (no co-resident services) and a longer per-run cap so that all 9 records flush to `runs.jsonl`. The script must write each record incrementally — this attempt lost 7 records because records were only flushed at the end of the loop. ## Files - `runs.jsonl` — the two records that completed (provenance field marks them as reconstructed verbatim from this stage's stdout after the script hit its cap before flush). - `probe_generate.json` — a minimal `num_predict:4` response used to confirm the endpoint was live and fast (HTTP 200, 0.76 s) after an earlier request had stalled on model load. - `raw-records-manifest.json` — file sizes and sha256 for every artifact, plus failed fetches.